Machine Learning Assisted Differentiation of Low Acuity Patients at Dispatch (MADLAD)

Machine Learning Assisted Differentiation of Low Acuity Patients at Dispatch: A Randomized Controlled Trial

BACKGROUND: At Emergency Medical Dispatch (EMD) centers, Resource Constrained Situations (RCS) where there are more callers requiring an ambulance than there are available ambulances are common. At the EMD centers in Uppsala and Västmanland, patients experiencing these situations are typically assigned a low-priority response, are often elderly, and have non-specific symptoms. Machine learning techniques offer a promising but largely untested approach to assessing risks among these patients. OBJECTIVES: To establish whether the provision of machine learning-based risk scores improves the ability of dispatchers to identify patients at high risk for deterioration in RCS. DESIGN: Multi-centre, parallel-grouped, randomized, analyst-blinded trial. POPULATION: Adult patients contacting the national emergency line (112), assessed by a dispatch nurse in Uppsala or Västmanland as requiring a low-priority ambulance response, and experiencing an RCS. OUTCOMES: Primary: 1. Proportion of RCS where the first available ambulance was dispatched to the patient with the highest National Early Warning Score (NEWS) score Secondary: Difference in composite risk score consisting of ambulance interventions, emergent transport, hospital admission, intensive care, and mortality between patients receiving immediate vs. delayed ambulance response during RCS. Difference in NEWS between patients receiving immediate vs. delayed ambulance response during RCS. INTERVENTION: A machine learning model will estimate the risk associated with each patient involved in the RCS, and propose a patient to receive the available ambulance. In the intervention arm only, the assessment will be displayed in a user interface integrated into the dispatching system. TRIAL SIZE: 1500 RCS each consisting of multiple patients randomized 1:1 to control and intervention arms

Groups of patients experiencing a resource constrained situation randomized 1:1 at time of inclusion to control/intervention arms

Analyst masked to treatment group allocation in final analysis. Outcomes extracted algorithmically from databases.

Calculation of risk assessment score by machine learning algorithm and display of risk assessment information to dispatch nurses. Staff encouraged but not required to comply with suggested ranking.

A machine learning algorithm (Gradient boosting) applied to structured data collected in the Alitis Clinical Decision Support system, patient demographics, and free-text notes.

Proportion of RCS where the first available ambulance was dispatched to the patient with the highest National Early Warning Score (NEWS).

NEWS is a widely used and well-validated scoring algorithm based on objective patient vital signs, which are not causally dependent on the outcomes used to train the machine learning models investigated in this study. NEWS values will be based on the first set of vital signs obtained by ambulance nurses upon making contact with the patient. NEWS is measured on a 0-21 scale, with higher values corresponding to patients at higher risk for deterioration.

This measure investigates a composite score consisting of the outcomes used to train the machine learning models. The composite score is generated by identifying the following patient outcomes and assigning the corresponding weights: Abnormal intitial Arway/Breathing/Circulation findings by ambulance crew (4) Emergent (lights and sirens) transport to the hospital (2) Provision of prehospital interventions (1) Admission to in-patient care or mortality within 30 days (1) This results in a score from 0-8, with higher scores representing more

Difference in National Early Warning Score (NEWS) between patients with immediate vs. delayed response.

Inclusion Criteria: Identification of a resource constrained situation by ambulance director (i.e., 2 or more patients awaiting an ambulance response) Assigned priority 2A or 2B (Low-priority ambulance response) by dispatch nurse call-taker Valid Swedish personal identification number collected at dispatch Age >= 18 years Exclusion Criteria: Relevant calls received more than 30 minutes apart Logistical factors (eg. the patients' geographical locations) affect the ambulance assignment decision On scene risk factors (eg. a patient is outdoors and risks hypothermia) or risk mitigators (eg. healthcare staff already on-scene with a patient) affect the ambulance assignment decision

Spangler D, Hermansson T, Smekal D, Blomberg H. A validation of machine learning-based risk scores in the prehospital setting. PLoS One. 2019 Dec 13;14(12):e0226518. doi: 10.1371/journal.pone.0226518. eCollection 2019.

Spangler D, Edmark L, Winblad U, Collden-Benneck J, Borg H, Blomberg H. Using trigger tools to identify triage errors by ambulance dispatch nurses in Sweden: an observational study. BMJ Open. 2020 Mar 19;10(3):e035004. doi: 10.1136/bmjopen-2019-035004.